Deploying a financial services AI copilot: how we cut call time 27% and raised first-contact resolution to 81%

In the fall, we spent time on the floor of a private bank's call center with a headset on. During the third call we listened to, a customer asked about early payoff on installment plans: whether it triggered a penalty and how the penalty was calculated. The agent muted the call, opened tabs in three different systems, pulled up product terms, then switched to an internal wiki to find compliant language. The whole process took a minute and fifty seconds. The customer on the line was getting impatient.

This wasn't unusual. Spend an afternoon tracking call patterns, and you see agents look up information two to three times per call. Each lookup consumes more than 30% of the total talk time. The answers they find aren't always right. Compliance language changes every six months, but outdated versions stay in personal notes.

The problem isn't agent effort. Agents have to be search engines, compliance advisors, and customer support all at once.

What this financial services AI copilot actually solves

A financial services AI copilot is a sidebar tool that shows relevant answers, regulatory requirements, and compliant language to the agent's screen in real-time. It doesn't replace the agent; it removes the need to search through multiple systems so the agent can focus on the customer.

Financial services has unique constraints. One wrong statement like "this product guarantees returns" can result in regulatory fines. A copilot in this space can't simply plug in a language model and launch. Every answer must be accurate, and every piece of guidance must be traceable to source documentation.

This is why we built it around RAG, retrieval-augmented generation, with two separate knowledge bases.

Why RAG instead of just asking the model

Asking a model directly creates hallucination problems. It might give you a plausible number for a credit card's APR that's actually from three years ago. In financial services, this kind of error can't happen.

RAG retrieves the most relevant excerpts from the company's knowledge base when an agent asks a question: terms, rate cards, approved language. The model then composes an answer based only on that content. The model handles language; the knowledge base provides facts.

Two knowledge bases typically feed into the system. The first is the product knowledge base: rates, terms, account-opening workflows, dispute resolution procedures. The second is the compliance knowledge base: regulator-mandated disclosure requirements, prohibited language, standard scripts. These are maintained separately because they're updated on different schedules by different teams. Product owns rates; compliance owns wording. Both feed into the system in real-time.

Every response from the copilot includes a citation showing which document and version it came from. This audit trail is essential for passing internal audit and regulatory review. During a demo, no one asks about sources. After launch, auditors ask about this first thing.

Three months in: what the numbers look like

The bank tracked two core metrics before and after: first-contact resolution (FCR) and average handle time (AHT). These metrics reveal call center performance clearly.

The FCR increase is straightforward. Agents no longer transfer calls or say they'll research and call back. They answer completely on the first call. AHT decreased because the dead time spent searching disappeared. Compliance language errors dropped because the copilot displayed the right phrases and prohibited language directly to agents. New hires didn't need to memorize compliance language.

The first month after launch, FCR only improved three percentage points. The team questioned whether it had failed. The problem wasn't the system. It was the knowledge base.

Some clauses were written in such dense legal language that the model could retrieve them but agents couldn't understand them. We spent two weeks with the compliance team and senior representatives rewriting answers for common questions into language agents could read directly from the screen. After this revision, performance improved significantly.

Knowledge base quality, not model intelligence, sets the ceiling.

The hard part isn't the technology. It's adoption.

Wiring RAG into the customer service platform and building a working version takes three to four weeks. Getting agents to use the system is harder.

Many systems with full feature sets get disabled by agents because the sidebar is slow or blocks the screen. Our engineers sat beside agents during calls, watching where they got stuck and noting when they ignored prompts. Senior reps don't read full answers; they want one key sentence and a source link. We redesigned the UI to show a one-sentence summary first, with details available on expand. Adoption increased from 41% to 85% through this iteration.

The one thing to remember At Tenten, we sequence financial copilot projects the same way: lock down the knowledge base and audit trail first, then have engineers work with agents to refine the tool until they use it, then watch FCR and AHT show the results. A polished demo doesn't matter. What counts is when agents actually click the sidebar tool and it gives them the right answer.